Vinicius Brasil

My coworker Iris isn't a person

My PM dropped a DM:

Hey Vini, users were pasting their API keys into our settings page and getting “invalid key”, but the keys were fine. Mind taking a look at this?

I knew what the issue was before I finished reading it. We weren’t trimming the trailing newline people copy along with the key. I knew, just as fast, what fixing it would cost me:

Leave the Slack thread, stash my current work, create a new branch, open the editor, run Claude Code, review the diff, merge the PR, deploy it to production, come back to Slack to say it’s done.

Nine steps for two lines of logic. So I did what any engineer would do: I decided it could wait until I finished my current task.

What I didn’t know is that just by acknowledging the new task I’d opened a loop my brain refused to close. There’s a name for it: the Zeigarnik effect.

The Zeigarnik effect: unfinished tasks pile up as cognitive load A head in profile stands in for the mind. Unfinished task-loops — small rings left open by a gap — drop into it one after another, each rattling while it stays open. A segmented meter labelled cognitive load climbs a notch with every loop. Then the loops all snap closed at once, go still, and dissolve, and the meter drops back to empty: the relief of clearing them. This is the Zeigarnik effect — unfinished work keeps its grip, and keeps adding load, until it is done. COGNITIVE LOAD OPEN LOOPS

This time I typed something different on Slack. @Iris, this is probably lacking a whitespace trim. Open a PR. A PR link appeared in the same thread. She’d found the right handler, trimmed it, and even added a test for the newline case. I skimmed the diff, hit merge, and went back to what I was doing. The loop closed within three minutes.

Iris is the AI agent we provisioned on Slack at CrewAI. She’s built on CrewAI itself and wired into our internal tooling. We tag @Iris and she can kick off crews, file Linear issues, run Claude Code, open PRs, send emails. None of this is new: Linear, GitHub, and Devin all let you hand work to an agent the way you’d assign it to a teammate, and Salesforce now calls Slack the interface for the agentic enterprise.

The top five types of work I handed Iris
FIG. The top five types of work I handed Iris

I’ll admit I was skeptical about agents autonomously writing to our codebases. And for real engineering work, that skepticism still holds. Hand Iris something complex and the cost of writing the code just becomes the cost of reviewing it, and I end up babysitting threads instead of writing software.

But that’s not what surprised me. What surprised me is that for the small stuff I context switch less, even for code.

Thanks, Iris. And please remember me when robots take over.